Method and system for training a neural network

Abstract

A method and a system for training a neural network. The method includes receiving, by a processing device, a training image, a reference label and a reference class activation map, the reference label and the reference class activation map associated with a corresponding unbiased image of the training image and generating, using the processing device, a class label and a class activation map based on the training image using the neural network. The method also includes calculating, using the processing device, a classification loss value based on differences between the reference label and the class label, and a class activation map loss value based on differences between the reference class activation map and the class activation map and updating, using the processing device, the neural network to minimise the classification loss value and the class activation map loss value to improve accuracy of the neural network in generation of the class label and the class activation map.

The invention claimed is: 
     
         1 . A method for training a neural network, the method comprising: receiving, by a processing device, a training image, a reference label and a reference class activation map, the reference label and the reference class activation map associated with a corresponding unbiased image of the training image, the reference class activation map indicating pixel-wise values corresponding to image regions that contributed to predicting the reference label; generating, using the processing device, a class label and a class activation map based on the training image using the neural network; calculating, using the processing device, a classification loss value based on differences between the reference label and the class label, and a class activation map loss value based on differences between the reference class activation map and the class activation map; and updating, using the processing device, the neural network to minimise the classification loss value and the class activation map loss value to improve accuracy of the neural network in generation of the class label and the class activation map. 
     
     
         2 . The method as claimed in  claim 1 , wherein updating the neural network to minimise the classification loss value and the class activation map loss value comprises:
 updating, using the processing device, the neural network to minimise a sum of the classification loss value and the class activation map loss value.   
     
     
         3 . The method as claimed in  claim 1 , further comprising:
 receiving, by the processing device, a reference bias value associated with the training image;   generating, using the processing device, a bias label based on the training image using the neural network;   calculating, using the processing device, a bias loss value based on differences between the reference bias value and the bias label; and   updating, using the processing device, the neural network to minimise the bias loss value.   
     
     
         4 . The method as claimed in  claim 1 , wherein receiving, by the processing device, the training image comprises generating, using the processing device, the training image based on a bias transformation of the corresponding unbiased image. 
     
     
         5 . The method as claimed in  claim 4 , wherein the bias transformation comprises one or more of a change in blur level, color temperature and day-night lighting of the unbiased image. 
     
     
         6 . The method as claimed in  claim 1 , wherein receiving the reference label and the reference class activation map comprises:
 receiving, by the processing device, the corresponding unbiased image; and   generating, using the processing device, the reference class activation map based on the corresponding unbiased image using a pre-trained neural network.   
     
     
         7 . A system for training a neural network, the system comprising: a processing device configured to: receive a training image, a reference label and a reference class activation map, the reference label and the reference class activation map associated with a corresponding unbiased image of the training image, the reference class activation map indicating pixel-wise values corresponding to image regions that contributed to predicting the reference label; generate a class label and a class activation map based on the training image using the neural network; calculate a classification loss value based on differences between the reference label and the class label, and a class activation map loss value based on differences between the reference class activation map and the class activation map; and update the neural network to minimise the classification loss value and the class activation map loss value to improve accuracy of the neural network in generation of the class label and the class activation map. 
     
     
         8 . The system as claimed in  claim 7 , wherein the processing device is configured to:
 update the neural network to minimise a sum of the classification loss value and the class activation map loss value.   
     
     
         9 . The system as claimed in  claim 7 , wherein the processing device is further configured to:
 receive a reference bias value associated with the training image;   generate a bias label based on the training image using the neural network;   calculate a bias loss value based on differences between the reference bias value and the bias label; and   update the neural network to minimise the bias loss value.   
     
     
         10 . The system as claimed in  claim 7 , wherein the processing device is configured to generate the training image based on a bias transformation of the corresponding unbiased image. 
     
     
         11 . The system as claimed in  claim 10 , wherein the bias transformation comprises one or more of a change in blur level, color temperature and day-night lighting of the unbiased image. 
     
     
         12 . The system as claimed in  claim 7 , wherein the processing device is configured to:
 receive the corresponding unbiased image; and   generate the reference class activation map based on the corresponding unbiased image using a pre-trained neural network.